Lesson 09 of 38 · AI 101 - 12 min
AI 101: Build Your First AI Roadmap
Turn a vague interest in AI into a defensible 30-day roadmap: choose one repeated, low-risk workflow, name the decision owner and the data it is allowed to touch, define the evidence that proves it worked, and set a stop-or-expand checkpoint - the exact discipline that separates the ~5% of AI pilots that deliver ROI from the ~95% that stall.
Renamed from the AI roadmap guide
A useful AI roadmap is not a list of tools to buy - it is a short decision plan. It names which repeated workflow will be improved first, why that workflow is worth touching, what data the tool is allowed to use, who owns the review, and the date the team will decide to stop, expand, or replace the approach. The reason to start this disciplined is not bureaucracy; it is survival odds. The independent data on AI projects is brutal: most pilots that begin as 'let's try this tool and see' never produce measurable value. The ones that do almost always share the same shape - one narrow, repeated, reviewable job with a named owner and a clear definition of success. This lesson teaches you to build that shape before you spend a dollar on a subscription.
Build your first AI roadmap, a 30-day plan
A branded walkthrough: pick one narrow, repeatable, low-risk workflow, assign ownership and data readiness, and run a disciplined 30-day pilot with a stop-or-expand checkpoint.
What to understand
- Start with a repeated job, not a general ambition to 'use AI'. A roadmap anchored to one recurring workflow is something you can scope, own, measure, and review; an ambition to 'adopt AI' is not.
- A first AI win should be frequent, visible, low-risk, and easy for a human to review. Frequency gives you fast feedback, visibility builds internal trust, low risk protects you while you learn, and reviewability means a person can catch errors before they reach a customer.
- Roadmaps fail when they skip ownership. MIT's 2025 study found the gap that kills pilots is organisational, not technical - nobody owned integrating the tool into a real workflow. Every workflow on your roadmap needs a single named person responsible for quality and adoption.
- Data readiness matters as much as model capability. Gartner reports the single largest cause of AI project failure is poor data quality; messy, scattered, or unlabelled inputs produce unreliable outputs no matter how strong the model is. Score the inputs you actually have, not the inputs you wish you had.
- Score before you choose. Rate each candidate workflow on impact, effort, data readiness, risk, and owner confidence - then pick the one with useful impact, moderate effort, low risk, and a confident owner. Do not let a high impact score override a high risk score on a first attempt.
- Define success as a business outcome, not 'the tool produced text'. Time saved, errors reduced, faster draft turnaround, cleaner handover, or better decision support - pick one measurable signal before you start so the 30-day review is a verdict, not a vibe.
- A 30-day roadmap is a learning loop with a built-in exit. Week one sets up, week two runs a draft, week three runs a reviewed version, week four is a checkpoint where you decide to expand, pause, or replace. The stop condition is part of the plan, not an admission of failure.
- Prefer back-office and internal-first workflows for your first win. The same MIT research found ROI showed up most reliably in back-office automation, not in the customer-facing sales and marketing tools where most budgets are spent - so a 'boring' internal workflow is often the smarter first roadmap target.
Deeper dive
Why the roadmap discipline exists: the 95% problem
In 2025 MIT's NANDA initiative published 'The GenAI Divide: State of AI in Business', based on 150 leader interviews, 350 employee surveys, and analysis of 300 public deployments. Its headline finding: roughly 95% of enterprise generative-AI pilots produced no measurable return, while only about 5% drove real revenue acceleration. The decisive variable was not model quality - the same frontier models were available to everyone. It was what MIT called the 'learning gap': winning organisations integrated AI into a specific workflow, with a specific owner, and learned from the results, while the rest ran disconnected experiments that never touched a real process. A roadmap is simply the deliberate version of what the 5% did by instinct. Every box on your one-pager - repeated workflow, named owner, allowed data, success metric, review checkpoint - is a direct countermeasure to a documented failure cause. You are not doing paperwork; you are buying yourself into the 5%.
Read the heatmap like an operator, not an optimist
The instinct is to pick the workflow with the highest impact score. That instinct is what fills the pilot graveyard. The heatmap exists to force a portfolio judgement across five dimensions at once. For a first win you are optimising for fast, safe proof - so you want high impact AND ready data AND low risk AND a confident owner, even if effort is only moderate. In the sample data, 'Weekly status pack' wins not because it is the flashiest but because it is the most balanced: strong impact (84), low effort (32), ready data (76), low risk (24), and a confident owner (82). 'Document summary' looks tempting but its risk (54) and weaker data readiness (58) make it a poor first attempt - the kind of workflow that quietly produces a confident-sounding wrong summary nobody catches. The rule: on your first roadmap, let risk and owner confidence act as veto columns. A high-impact idea with high risk and a hesitant owner is a second-quarter project, not a first win.
Data readiness is the silent kill switch
Gartner has repeatedly named poor data quality as the dominant cause of AI project failure, reporting that the large majority of stalled projects trace back to inputs that were scattered, inconsistent, or not allowed to be used. This is why the roadmap asks you to list the data you already have and are permitted to use before you choose a workflow - not after. A 'weekly status pack' usually scores well on data readiness because the inputs already exist in a known place (a project tracker, a few spreadsheets, last week's pack) and contain no regulated personal data. A 'candidate shortlist' might score higher on impact but lower on data readiness and far higher on risk, because the inputs are CVs containing personal information that carry privacy obligations. Score data readiness honestly and let it pull risk along with it: where the data is messy or sensitive, both scores should reflect it, and the workflow should drop down your list.
First-win workflow candidates, scored
A worked example of step 2: three of the heatmap candidates plus a deliberately risky contrast (candidate shortlist), scored 1-5 (5 = best) so you can see how the recommendation falls out. Reviewability and owner confidence act as veto columns for a first win - a high-impact idea with low reviewability and a hesitant owner is not a beginner workflow. Note the worked example scores Reviewability in place of Effort: for a first win, whether a person can actually check the output is usually the binding constraint, while effort shows up indirectly in every other column.
| Candidate workflow | Impact | Data ready | Risk (low = good) | Reviewability | Owner confidence | First-win verdict |
|---|---|---|---|---|---|---|
| Weekly status pack | 4 | 4 | 5 (low risk) | 5 | 5 | Strong first win - balanced, internal, easy to check |
| Public content brief | 4 | 5 | 4 | 4 | 4 | Good alternative - internal draft, human approves before publish |
| Customer email draft | 4 | 3 | 3 | 4 | 4 | Defer - customer-facing; add a human-approval gate first |
| Candidate shortlist | 5 | 2 | 1 (high risk) | 2 | 2 | Avoid for now - sensitive personal data, hard to review fairly |
Sources (as of June 2026): MIT NANDA - The GenAI Divide: State of AI in Business 2025 (via Fortune) · Gartner - AI projects stall ahead of ROI / data quality (Apr 2026)
First-win roadmap heatmap
Each candidate is scored 0-100 here so the heat colours have room to spread; in step 2 you will use a simpler 1-5 on the same five dimensions. Either way the read is identical: you want high impact, manageable effort, ready data, low risk, and a confident owner - not the single most impressive idea.
Step by step
List three to five repeated workflows
Write down three to five jobs that happen weekly or fortnightly. Use plain names: weekly status report, lead follow-up, policy summary, candidate shortlist, content brief. For each, note how often it happens and where its inputs already live (a folder, a spreadsheet, a tracker, an inbox).
HintIf a job happens only once, it is not your first roadmap candidate - frequency is what gives you fast feedback inside a 30-day window. You are done when each workflow has a frequency and an input location written next to it.
Score the candidates on five dimensions
Score each workflow 1-5 for impact, effort, data readiness, risk, and owner confidence (use the worked comparison table above as your model). Pick the candidate with useful impact, moderate effort, ready data, low risk, and a confident owner. Treat risk and owner confidence as veto columns.
HintDo not let impact override risk. A high-impact workflow with sensitive data or no clear way to review it is rarely a beginner first win - it is a second-quarter project. You are done when every candidate has five scores and one row clearly wins - or you have learned that none should go first yet.
On this screen
- 1What to notice. The winner is usually the most balanced row, not the highest single score - fast, safe, reviewable beats impressive-but-risky for a first win.
- 2Why it matters. MIT's 2025 study found ~95% of pilots failed not on model quality but on integration; a balanced, reviewable workflow is the kind that actually gets integrated.
Name the owner and the data boundary
Name one person who owns quality and adoption for the chosen workflow - not a committee. Then write what data the tool is allowed to use, what it may draft, what it must never access, and what a human must approve before anything leaves the team.
HintOwnership is the failure point the research keeps flagging. If you cannot name the single person responsible, the workflow is not ready to be on the roadmap. You are done when the boundary fits in four sentences a manager could sign.
On this screen
- 1What to notice. The data boundary belongs on the roadmap, not in someone's head - it is what keeps a 'first win' from becoming a privacy incident.
Write the 30-day plan and success metric
Lay out four weeks: week 1 setup, week 2 a draft run, week 3 a reviewed run, week 4 a decision checkpoint. Name the owner, inputs, review rule, and expected output. Then pick ONE measurable success metric - time saved, fewer errors, faster turnaround, cleaner handover, or better decision support.
HintThe roadmap should fit on one page. If it needs a deck, the first workflow is too broad - narrow it until a manager can read it in two minutes.
Set the checkpoint and the stop-or-expand rule
Decide in advance what evidence at day 30 means 'expand', what means 'adjust and retry', and what means 'stop'. Write the stop condition explicitly so the checkpoint is a verdict, not a debate. Capture the lesson learned either way.
HintA pre-decided stop condition is what makes this a learning loop instead of an experiment that drifts on forever consuming budget and attention. You are done when a colleague could run the day-30 checkpoint without you in the room.
On this screen
- 1Why it matters. Gartner predicts over 40% of agentic AI projects will be cancelled by end of 2027 for unclear value - a stop-or-expand rule means you cancel on purpose and early, not expensively and late.
Complete a one-page AI roadmap for one workflow. Score at least three candidates, then for the winner include the owner, inputs, allowed-data boundary, review checkpoint, single success metric, and explicit stop condition.
A one-page AI roadmap: first-win workflow, named owner, allowed inputs and data boundary, success metric, key risks, 30-day review checkpoint, and an explicit stop-or-expand decision rule.
Production prompt examples
ROLE: You are a pragmatic AI adoption advisor helping a non-technical business team pick its FIRST 30-day AI workflow. You optimise for fast, safe, provable wins - not the most ambitious idea. CONTEXT - about my team: - Team / function: [e.g. 6-person operations team in a B2B services firm] - Tools we already use: [e.g. Google Workspace, a CRM, a project tracker] - Data we already have and are allowed to use: [list it; flag anything containing personal or regulated data] - Review capacity: [who could check the AI output each week, and how much time they have] CANDIDATE WORKFLOWS (3-5 repeated jobs we do weekly or fortnightly): 1. [workflow] - happens [how often], inputs are [where they live] 2. [workflow] - ... 3. [workflow] - ... TASK: 1. Score each candidate 1-5 (5 = best) on: Impact, Data readiness, Risk (score it so LOW risk = 5), Reviewability, and Owner confidence. Show the scores in a table. 2. Recommend ONE workflow as the first 30-day win and explain in 3-4 sentences why it beats the others - explicitly referencing risk and reviewability, not just impact. 3. Draft a one-page 30-day plan for the winner: week 1 setup, week 2 draft run, week 3 reviewed run, week 4 decision checkpoint. Name the owner, the allowed inputs, the human-review rule, the single success metric, and an explicit stop condition. CONSTRAINTS: - Do NOT recommend buying any new tool unless an existing tool genuinely cannot do the job - and if so, say exactly why. - Treat any workflow with sensitive personal data or a customer-facing output as higher risk and require a human-approval gate before anything leaves the team. - Keep the whole answer under one page. Use plain business language a manager who never opens the tool could follow. OUTPUT FORMAT: 1. Scoring table. 2. Recommendation (3-4 sentences). 3. One-page 30-day plan with owner, inputs, review rule, success metric, stop condition.
- The ROLE line ('fast, safe, provable wins - not the most ambitious idea') is what stops the model from defaulting to the flashy high-impact-high-risk pick that fills the pilot graveyard.
- Asking the team to list 'data we already have and are allowed to use' up front bakes in the Gartner data-readiness reality before any workflow is chosen.
- Scoring risk so that LOW risk = 5 keeps every column pointing the same direction, so the highest total is genuinely the safest-best first win.
- Forcing the recommendation to 'explicitly reference risk and reviewability, not just impact' is the single instruction that most changes the answer for the better.
- The 'do not recommend buying a tool unless an existing one cannot do it' constraint directly counters the most common and expensive mistake: buying a subscription before naming the job.
- Requiring an explicit stop condition turns the 30-day plan into a real decision loop rather than an open-ended experiment that quietly never ends.
Common mistakes to avoid
- Starting with a tool subscription before naming the repeated workflow - the most common and expensive first mistake, and the one MIT's data ties directly to failed pilots.
- Choosing a sensitive or customer-facing process before the team has a review habit, instead of proving the loop on a low-risk internal workflow first.
- Skipping the owner and hoping the tool will create adoption by itself - the 'learning gap' that the research names as the decisive failure cause.
- Ignoring data readiness and picking the highest-impact idea even when its inputs are messy or off-limits - Gartner's top cause of AI project failure.
- Defining success as 'the tool produced text' rather than a measurable business outcome, so the 30-day checkpoint has no verdict.
- Calling a wishlist a roadmap. A roadmap needs sequence, an owner, a success metric, and a stop-or-expand checkpoint.
Source conflicts to review
- The '95% of GenAI pilots fail' headline is from MIT NANDA's 2025 study and counts pilots with no measurable P&L impact; some commentators argue the figure overstates failure because of how 'pilot' and 'ROI' are defined. Use it as directional support for starting narrow and measurable, not as a precise statistic.
- AI project failure-rate estimates vary widely by source and definition (Gartner, RAND, and MIT report different numbers from different samples and methods). The consistent signal across all of them - unclear value, weak ownership, and poor data readiness - is what this lesson is built to counter, regardless of the exact percentage.
Key terms
- First-win workflow
- A narrow, repeated task chosen to prove AI value quickly without creating major risk - the anchor of the whole roadmap.
- Owner confidence
- How confident the single responsible person is that they can review the output and drive adoption of the workflow.
- Data readiness
- Whether the inputs a workflow needs already exist, are consistent, and are allowed to be used - the most common silent cause of AI project failure.
- Review checkpoint
- A pre-planned day-30 decision point where the team expands, adjusts, or stops the pilot based on evidence.
- Stop condition
- The agreed result that means 'do not continue' - written in advance so the checkpoint is a verdict rather than a debate.
- Learning gap
- MIT's term for the organisational failure to integrate an AI tool into a real workflow, owner, and feedback loop - the main reason most pilots stall.
- ROI
- Return on investment - whether the time and money put into the pilot came back as measurable value.
- P&L impact
- An effect that shows up in the profit-and-loss numbers - real revenue gained or cost removed, not just activity.
- Pilot
- A deliberately small first deployment run to learn whether an approach works before committing more budget.
Resources
Checkpoint